A 24/7 emergency triaging phone call application that connects injured or endangered individuals with AI agents for situation assessment and guidance. The system persists caller memory across sessions and provides a real-time dashboard for monitoring active calls.

• Frontend: React + TypeScript + Mantine UI
• Backend: Node.js + Express + TypeScript
• Database: PostgreSQL with emergency call schemas
• Voice API: VAPI integration for phone calls
• AI Agents: Letta (formerly MemGPT) for persistent memory
• Real-time: WebSocket connections for dashboard updates
• Hosting: Vercel (frontend) + Railway/Render (backend)

┌─────────────────┐    ┌──────────────────┐    ┌─────────────────┐
│   Emergency     │    │   VAPI Voice     │    │   Dashboard     │
│   Caller        │◄──►│   Assistant      │◄──►│   Operators     │
└─────────────────┘    └──────────────────┘    └─────────────────┘
                              │                          │
                              ▼                          ▼
┌─────────────────┐    ┌──────────────────┐    ┌─────────────────┐
│   Letta AI      │◄──►│   Backend API    │◄──►│   Real-time     │
│   Memory Agent  │    │   (Node.js)      │    │   WebSocket     │
└─────────────────┘    └──────────────────┘    └─────────────────┘
                              │
                              ▼
                    ┌──────────────────┐
                    │   PostgreSQL     │
                    │   Database       │
                    └──────────────────┘

• Node.js 18+ and npm
• PostgreSQL 14+
• VAPI account and API key
• Letta account and API key

# Clone the repository
git clone <repository-url>
cd emergency-triage

# Install backend dependencies
cd backend
npm install

# Install frontend dependencies
cd ../frontend
npm install

Create .env files for both backend and frontend:

Backend .env:

DATABASE_URL=postgresql://username:password@localhost:5432/emergency_triage
VAPI_API_KEY=your_vapi_api_key
VAPI_PHONE_NUMBER=your_emergency_number
LETTA_API_KEY=your_letta_api_key
JWT_SECRET=your_jwt_secret
PORT=3001
FRONTEND_URL=http://localhost:3000

Frontend .env:

REACT_APP_API_BASE_URL=http://localhost:3001
REACT_APP_WS_URL=http://localhost:3001

# Create PostgreSQL database
createdb emergency_triage

# Run setup and test script
npx ts-node setup-and-test.ts

# Terminal 1: Start backend
cd backend
npm run dev

# Terminal 2: Start frontend
cd frontend
npm start

Open http://localhost:3000 to access the real-time emergency dashboard.

• Instant Response: < 3 second call answering
• AI Triage: Automated situation assessment
• Multi-language: English, Spanish, French support, etc
• Severity Classification: CRITICAL, HIGH, MEDIUM, LOW levels
• Real-time Transcription: Live speech-to-text processing

• Caller Profiles: Automatic identification via phone number
• Medical History: Persistent health condition tracking
• Emergency History: Previous call outcomes and patterns
• Location Data: GPS coordinates and address storage
• Risk Assessment: Dynamic risk profile calculation

• Live Call Monitoring: Active call status and transcriptions
• Emergency Alerts: Critical situation notifications
• Caller Context: Historical data and risk profiles
• Escalation Controls: Human operator handoff
• Analytics: Call volume, response times, outcomes

• Auto-escalation: Based on severity and caller history
• Human Operators: Specialized emergency response staff
• Emergency Services: Direct 911/emergency services dispatch
• Supervisor Alerts: Complex situation management

# Calls Management
POST   /api/calls/initiate        # Start emergency call
GET    /api/calls/active          # Get active calls
GET    /api/calls/history         # Get call history
POST   /api/calls/escalate        # Escalate to human operator
POST   /api/calls/webhook/vapi    # VAPI webhook handler

# User & Memory Management
GET    /api/users/:id             # Get user profile
POST   /api/users                 # Create user profile
GET    /api/users/:id/memory      # Get caller memory
PUT    /api/users/:id/memory      # Update caller memory

# Dashboard & Analytics
GET    /api/dashboard/stats       # Dashboard statistics
GET    /api/dashboard/active-calls # Active calls for dashboard
GET    /api/dashboard/recent-calls # Recent call history

// Client-side event listeners
socket.on("newCall", (call) => {
  /* Handle new emergency call */
});
socket.on("callUpdate", (event) => {
  /* Handle call status update */
});
socket.on("callEscalated", (data) => {
  /* Handle escalation */
});
socket.on("emergencyAlert", (alert) => {
  /* Handle critical alert */
});

-- User profiles and caller information
users (id, phone_number, risk_profile, preferred_language, created_at)

-- Emergency call records
calls (id, user_id, emergency_type, severity_level, status, transcription, outcome, created_at)

-- Persistent caller memory
memory_records (id, user_id, memory_type, content, created_at, updated_at)

-- Escalation tracking
escalation_log (id, call_id, escalation_type, reason, created_at)

-- Emergency services dispatch
emergency_dispatch_log (id, call_id, service_type, dispatch_time, status)

The system uses Letta AI agents for persistent caller memory:

// Caller memory structure
interface CallerMemory {
  userId: string;
  phoneNumber: string;
  medicalHistory: string[];
  previousEmergencies: EmergencyRecord[];
  location: LocationData;
  emergencyContacts: Contact[];
  riskProfile: "LOW" | "MEDIUM" | "HIGH" | "CRITICAL";
  lastCallSummary: string;
}

1. Immediate Assessment: Severity classification within 30 seconds
2. Contextual Analysis: Historical pattern recognition
3. Escalation Logic: Automated decision tree for human handoff
4. Memory Updates: Real-time learning from each interaction

// Emergency triage assistant settings
{
  "name": "Emergency Triage Assistant",
  "model": "gpt-4",
  "voice": "professional-calm",
  "systemMessage": "Emergency triage AI with medical protocols...",
  "functions": ["escalate_to_human", "dispatch_emergency_services"],
  "recordingEnabled": true,
  "maxDurationSeconds": 1800
}

1. Incoming Call: VAPI answers with emergency greeting
2. Situation Assessment: Structured triage questions
3. Memory Retrieval: Access caller's historical context
4. AI Analysis: Severity determination and guidance
5. Escalation Decision: Human/emergency services if needed
6. Memory Update: Store call outcomes and learnings

• CRITICAL: Life-threatening (heart attack, stroke, severe bleeding)
• HIGH: Serious but stable (broken bones, difficulty breathing)
• MEDIUM: Concerning but manageable (moderate injury, infection)
• LOW: Minor issues (small cuts, mild symptoms)

• Critical severity assessment
• High-risk caller with concerning symptoms
• Extended call duration (>10 minutes)
• Caller becomes unresponsive
• Specific keywords ("can't breathe", "unconscious")

• Active Calls: Real-time emergency call count
• Response Times: Average time to answer and assess
• Escalation Rate: Percentage requiring human intervention
• Outcome Tracking: Call resolution and follow-up
• Risk Distribution: Caller risk profile analytics

• High call volume detection
• Extended response times
• Service availability issues
• Critical emergency patterns

• HIPAA-compliant medical information handling
• Encrypted database storage
• Secure API authentication
• Audit logging for all actions

• Override capabilities for supervisors
• Emergency service direct dispatch
• Call recording and transcription storage
• Regulatory compliance reporting

# Frontend (Vercel)
vercel --prod

# Backend (Railway)
railway login
railway deploy

# Database (Railway PostgreSQL)
railway add postgresql

Ensure all production environment variables are configured:

• Database connection strings
• API keys for VAPI and Letta
• JWT secrets and security tokens
• Emergency service integration credentials

Run the comprehensive test suite:

# Setup database and run all tests
npx ts-node setup-and-test.ts

# Individual test components
npm run test:emergency-flow
npm run test:dashboard
npm run test:escalation

• Call Answer: < 3 seconds
• Initial Assessment: < 30 seconds
• Escalation: < 60 seconds for critical cases
• Memory Retrieval: < 1 second

• Database connection pooling
• WebSocket connection management
• Caching for frequent queries
• Load balancing for high call volumes

• Video call capability for visual assessment
• Integration with local emergency services
• Advanced AI diagnosis and treatment recommendations
• Mobile app for emergency contacts
• Geographic routing to local resources
• Multi-language support expansion
• Wearable device integration
• Predictive emergency analysis

For technical support or questions about the Emergency Triage System:

• Documentation: See inline code comments and API docs
• Issues: Report bugs via the issue tracker
• Emergency Protocol Questions: Consult medical professionals
• System Status: Monitor dashboard health indicators

⚠️ IMPORTANT: This system is designed to assist with emergency triage but should never replace professional medical advice or emergency services. Always call 911 or local emergency services for life-threatening situations.